Tencent, the Chinese technology conglomerate, officially released 'Hy3' on July 6, 2026, a foundational model built specifically for reasoning tasks and AI agent deployment. This version improves upon the performance, stability, and cost-efficiency of the preview model released in April.

Hy3 is a Mixture-of-Experts (MoE) architecture featuring 295B total parameters, 21B active parameters, and a 3.8B MTP layer. It supports a context window of up to 256K tokens and utilizes a configuration that activates the top 8 of 192 total experts. Benchmark results indicate improved performance in both reasoning and agent-related tasks.

The model is already integrated into Tencent products, including WorkBuddy, Yuanbao, Marvis, and ima. It also powers the official Weixin/WeChat account assistant and AI features within the game 'Path of Exile: Advent'. For developers, the model is available via the Apache 2.0 license on platforms such as Hugging Face and ModelScope.

Access is provided through Tencent Cloud TokenHub, with pricing set at 1 CNY per million input tokens, 4 CNY for output tokens, and 0.25 CNY for cache hits. Future updates will expand integration to external platforms including OpenRouter, Cline, and Cherry Studio.

Tencent, the Chinese technology conglomerate, officially released 'Hy3' on July 6, 2026, a foundational model built specifically for reasoning tasks and AI agent deployment. This version improves upon the performance, stability, and cost-efficiency of the preview model released in April.

Hy3 is a Mixture-of-Experts (MoE) architecture featuring 295B total parameters, 21B active parameters, and a 3.8B MTP layer. It supports a context window of up to 256K tokens and utilizes a configuration that activates the top 8 of 192 total experts. Benchmark results indicate improved performance in both reasoning and agent-related tasks.

The model is already integrated into Tencent products, including WorkBuddy, Yuanbao, Marvis, and ima. It also powers the official Weixin/WeChat account assistant and AI features within the game 'Path of Exile: Advent'. For developers, the model is available via the Apache 2.0 license on platforms such as Hugging Face and ModelScope.

Access is provided through Tencent Cloud TokenHub, with pricing set at 1 CNY per million input tokens, 4 CNY for output tokens, and 0.25 CNY for cache hits. Future updates will expand integration to external platforms including OpenRouter, Cline, and Cherry Studio.